Comparison of passive microwave brightness temperature prediction sensitivities over snow-covered land in North America using machine learning algorithms and the Advanced Microwave Scanning Radiometer

Recent studies showed that machine learning (ML) algorithms (e.g., artificial neural network (ANN) and support vector machine (SVM)) reasonably reproduce passive microwave brightness temperature observations over snow-covered land as measured by the Advanced Microwave Scanning Radiometer (AMSR-E). However, these studies did not explore the sensitivities of the ML algorithms relative to ML inputs in order to determine the behavior and performance of each algorithm. In this current study, normalized sensitivity coefficients are computed to diagnose ML performance as a function of time and space. The results showed that when using the ANN, approximately 20% of locations across North America are relatively sensitive to snow water equivalent (SWE). However, more than 65% of locations in the SVM-based brightness temperature (Tb) estimates are sensitive relative to perturbations in SWE at all frequency and polarization combinations explored in this study. Further, the SVM-based results suggest the algorithm is sensitive in both shallow and deep SWE, SWE with and without overlying forest canopy, and during both the snow accumulation and snow ablation seasons. Therefore, these findings suggest that compared with the ANN, the SVM could potentially serve as a more efficient and effective measurement model operator within a Tb data assimilation framework for the purpose of improving SWE estimates across regional- and continental-scales.